Skip to content

armgit / dify

Public Imported from GitHub
0

Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack.

8.2 KB
Raw Blame
import logging
import time

import click
from celery import shared_task
from sqlalchemy import delete, select

from configs import dify_config
from core.app.apps.pipeline.pipeline_generator import PipelineGenerator
from core.db.session_factory import get_session_maker, session_factory
from core.rag.index_processor.index_processor import IndexProcessorFactory
from enums import DeploymentEdition
from extensions.application_services.data_sources import build_data_source_credentials
from extensions.ext_redis import redis_client
from libs.datetime_utils import naive_utc_now
from models import Account, Tenant
from models.dataset import Dataset, Document, DocumentSegment
from models.enums import IndexingStatus
from repositories.knowledge.document_repository import SQLAlchemyDocumentRepository
from services.feature_service import FeatureService
from services.knowledge.indexing.adapters.execution import build_document_indexing_service
from services.knowledge.resource_scope import DatasetRef
from services.rag_pipeline.rag_pipeline import RagPipelineService
from services.workflow_run_agg import WorkflowRunAgg

logger = logging.getLogger(__name__)


@shared_task(queue="dataset")
def retry_document_indexing_task(dataset_id: str, document_ids: list[str], user_id: str):
    """
    Async process document
    :param dataset_id:
    :param document_ids:
    :param user_id:

    Usage: retry_document_indexing_task.delay(dataset_id, document_ids, user_id)
    """
    start_at = time.perf_counter()
    with session_factory.create_session() as session:
        try:
            dataset = session.scalar(select(Dataset).where(Dataset.id == dataset_id).limit(1))
            if not dataset:
                logger.info(click.style(f"Dataset not found: {dataset_id}", fg="red"))
                return
            user = session.scalar(select(Account).where(Account.id == user_id).limit(1))
            if not user:
                logger.info(click.style(f"User not found: {user_id}", fg="red"))
                return
            tenant = session.scalar(select(Tenant).where(Tenant.id == dataset.tenant_id).limit(1))
            if not tenant:
                raise ValueError("Tenant not found")
            user.set_current_tenant_with_session(tenant, session=session)

            for document_id in document_ids:
                retry_indexing_cache_key = f"document_{document_id}_is_retried"
                # check document limit
                if dify_config.DEPLOYMENT_EDITION == DeploymentEdition.CLOUD:
                    features = FeatureService.get_features(tenant.id)
                    try:
                        vector_space = features.vector_space
                        assert vector_space is not None
                        if 0 < vector_space.limit <= vector_space.size:
                            raise ValueError(
                                "Your total number of documents plus the number of uploads have over the limit of "
                                "your subscription."
                            )
                    except Exception as e:
                        document = session.scalar(
                            select(Document)
                            .where(Document.id == document_id, Document.dataset_id == dataset_id)
                            .limit(1)
                        )
                        if document:
                            document.indexing_status = IndexingStatus.ERROR
                            document.error = str(e)
                            document.stopped_at = naive_utc_now()
                            session.add(document)
                            session.commit()
                        redis_client.delete(retry_indexing_cache_key)
                        return

                logger.info(click.style(f"Start retry document: {document_id}", fg="green"))
                document = session.scalar(
                    select(Document).where(Document.id == document_id, Document.dataset_id == dataset_id).limit(1)
                )
                if not document:
                    logger.info(click.style(f"Document not found: {document_id}", fg="yellow"))
                    return
                try:
                    # clean old data
                    index_processor = IndexProcessorFactory(document.doc_form).init_index_processor()

                    segments = session.scalars(
                        select(DocumentSegment).where(DocumentSegment.document_id == document_id)
                    ).all()
                    if segments:
                        index_node_ids = [segment.index_node_id for segment in segments if segment.index_node_id]
                        # delete from vector index
                        index_processor.clean(
                            dataset,
                            index_node_ids,
                            with_keywords=True,
                            delete_child_chunks=True,
                            session=session,
                        )

                    segment_ids = [segment.id for segment in segments]
                    if segment_ids:
                        segment_delete_stmt = delete(DocumentSegment).where(DocumentSegment.id.in_(segment_ids))
                        session.execute(segment_delete_stmt)
                    session.commit()

                    document.indexing_status = IndexingStatus.PARSING
                    document.processing_started_at = naive_utc_now()
                    session.add(document)
                    document_ref = DatasetRef(document.tenant_id, document.dataset_id).document(document.id)
                    runtime_mode = dataset.runtime_mode
                    # Persist preparation before handing off to the selected execution service.
                    session.commit()

                    if runtime_mode == "rag_pipeline":
                        with session_factory.create_session() as rag_session:
                            rag_pipeline_service = RagPipelineService(rag_session)
                            rag_pipeline_service.retry_error_document(
                                dataset,
                                document,
                                user,
                                generator=PipelineGenerator(
                                    execution_driver=WorkflowRunAgg.run,
                                    documents=SQLAlchemyDocumentRepository(
                                        session_factory=session_factory.get_session_maker()
                                    ),
                                    datasource_providers=build_data_source_credentials(
                                        database_client=get_session_maker()
                                    ).providers,
                                ),
                            )
                    else:
                        indexing_service = build_document_indexing_service(
                            session_factory=session_factory.get_session_maker(), enforce_vector_space_admission=True
                        )
                        indexing_service.run([document_ref])
                    redis_client.delete(retry_indexing_cache_key)
                except Exception as ex:
                    session.rollback()
                    document = session.scalar(
                        select(Document).where(Document.id == document_id, Document.dataset_id == dataset_id).limit(1)
                    )
                    if document:
                        document.indexing_status = IndexingStatus.ERROR
                        document.error = str(ex)
                        document.stopped_at = naive_utc_now()
                        session.add(document)
                        session.commit()
                    logger.info(click.style(str(ex), fg="yellow"))
                    redis_client.delete(retry_indexing_cache_key)
                    logger.exception("retry_document_indexing_task failed, document_id: %s", document_id)
            end_at = time.perf_counter()
            logger.info(click.style(f"Retry dataset: {dataset_id} latency: {end_at - start_at}", fg="green"))
        except Exception as e:
            logger.exception(
                "retry_document_indexing_task failed, dataset_id: %s, document_ids: %s", dataset_id, document_ids
            )
            raise e